2 * Copyright (c) 2016 Martin Storsjo
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License along
17 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 #include "libavcodec/avcodec.h"
24 #include "libavcodec/vp8dsp.h"
26 #include "libavutil/common.h"
27 #include "libavutil/intreadwrite.h"
28 #include "libavutil/mem_internal.h"
32 #define PIXEL_STRIDE 16
34 #define randomize_buffers(src, dst, stride, coef) \
37 for (y = 0; y < 4; y++) { \
38 AV_WN32A((src) + y * (stride), rnd()); \
39 AV_WN32A((dst) + y * (stride), rnd()); \
40 for (x = 0; x < 4; x++) \
41 (coef)[y * 4 + x] = (src)[y * (stride) + x] - \
42 (dst)[y * (stride) + x]; \
46 static void dct4x4(int16_t *coef)
49 for (i = 0; i < 4; i++) {
50 const int a1 = (coef[i*4 + 0] + coef[i*4 + 3]) * 8;
51 const int b1 = (coef[i*4 + 1] + coef[i*4 + 2]) * 8;
52 const int c1 = (coef[i*4 + 1] - coef[i*4 + 2]) * 8;
53 const int d1 = (coef[i*4 + 0] - coef[i*4 + 3]) * 8;
54 coef[i*4 + 0] = a1 + b1;
55 coef[i*4 + 1] = (c1 * 2217 + d1 * 5352 + 14500) >> 12;
56 coef[i*4 + 2] = a1 - b1;
57 coef[i*4 + 3] = (d1 * 2217 - c1 * 5352 + 7500) >> 12;
59 for (i = 0; i < 4; i++) {
60 const int a1 = coef[i + 0*4] + coef[i + 3*4];
61 const int b1 = coef[i + 1*4] + coef[i + 2*4];
62 const int c1 = coef[i + 1*4] - coef[i + 2*4];
63 const int d1 = coef[i + 0*4] - coef[i + 3*4];
64 coef[i + 0*4] = (a1 + b1 + 7) >> 4;
65 coef[i + 1*4] = ((c1 * 2217 + d1 * 5352 + 12000) >> 16) + !!d1;
66 coef[i + 2*4] = (a1 - b1 + 7) >> 4;
67 coef[i + 3*4] = (d1 * 2217 - c1 * 5352 + 51000) >> 16;
71 static void wht4x4(int16_t *coef)
74 for (i = 0; i < 4; i++) {
75 int a1 = coef[0 * 4 + i];
76 int b1 = coef[1 * 4 + i];
77 int c1 = coef[2 * 4 + i];
78 int d1 = coef[3 * 4 + i];
92 for (i = 0; i < 4; i++) {
93 int a1 = coef[i * 4 + 0];
94 int b1 = coef[i * 4 + 1];
95 int c1 = coef[i * 4 + 2];
96 int d1 = coef[i * 4 + 3];
105 coef[i * 4 + 0] = a1 * 2;
106 coef[i * 4 + 1] = c1 * 2;
107 coef[i * 4 + 2] = d1 * 2;
108 coef[i * 4 + 3] = b1 * 2;
112 static void check_idct(void)
114 LOCAL_ALIGNED_16(uint8_t, src, [4 * 4]);
115 LOCAL_ALIGNED_16(uint8_t, dst, [4 * 4]);
116 LOCAL_ALIGNED_16(uint8_t, dst0, [4 * 4]);
117 LOCAL_ALIGNED_16(uint8_t, dst1, [4 * 4]);
118 LOCAL_ALIGNED_16(int16_t, coef, [4 * 4]);
119 LOCAL_ALIGNED_16(int16_t, subcoef0, [4 * 4]);
120 LOCAL_ALIGNED_16(int16_t, subcoef1, [4 * 4]);
123 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, ptrdiff_t stride);
126 randomize_buffers(src, dst, 4, coef);
130 for (dc = 0; dc <= 1; dc++) {
131 void (*idct)(uint8_t *, int16_t *, ptrdiff_t) = dc ? d.vp8_idct_dc_add : d.vp8_idct_add;
133 if (check_func(idct, "vp8_idct_%sadd", dc ? "dc_" : "")) {
135 memset(subcoef0, 0, 4 * 4 * sizeof(int16_t));
136 subcoef0[0] = coef[0];
138 memcpy(subcoef0, coef, 4 * 4 * sizeof(int16_t));
140 memcpy(dst0, dst, 4 * 4);
141 memcpy(dst1, dst, 4 * 4);
142 memcpy(subcoef1, subcoef0, 4 * 4 * sizeof(int16_t));
143 // Note, this uses a pixel stride of 4, even though the real decoder uses a stride as a
144 // multiple of 16. If optimizations want to take advantage of that, this test needs to be
145 // updated to make it more like the h264dsp tests.
146 call_ref(dst0, subcoef0, 4);
147 call_new(dst1, subcoef1, 4);
148 if (memcmp(dst0, dst1, 4 * 4) ||
149 memcmp(subcoef0, subcoef1, 4 * 4 * sizeof(int16_t)))
152 bench_new(dst1, subcoef1, 4);
157 static void check_idct_dc4(void)
159 LOCAL_ALIGNED_16(uint8_t, src, [4 * 4 * 4]);
160 LOCAL_ALIGNED_16(uint8_t, dst, [4 * 4 * 4]);
161 LOCAL_ALIGNED_16(uint8_t, dst0, [4 * 4 * 4]);
162 LOCAL_ALIGNED_16(uint8_t, dst1, [4 * 4 * 4]);
163 LOCAL_ALIGNED_16(int16_t, coef, [4], [4 * 4]);
164 LOCAL_ALIGNED_16(int16_t, subcoef0, [4], [4 * 4]);
165 LOCAL_ALIGNED_16(int16_t, subcoef1, [4], [4 * 4]);
168 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t block[4][16], ptrdiff_t stride);
172 for (chroma = 0; chroma <= 1; chroma++) {
173 void (*idct4dc)(uint8_t *, int16_t[4][16], ptrdiff_t) = chroma ? d.vp8_idct_dc_add4uv : d.vp8_idct_dc_add4y;
174 if (check_func(idct4dc, "vp8_idct_dc_add4%s", chroma ? "uv" : "y")) {
175 ptrdiff_t stride = chroma ? 8 : 16;
176 int w = chroma ? 2 : 4;
177 for (i = 0; i < 4; i++) {
178 int blockx = 4 * (i % w);
179 int blocky = 4 * (i / w);
180 randomize_buffers(src + stride * blocky + blockx, dst + stride * blocky + blockx, stride, coef[i]);
182 memset(&coef[i][1], 0, 15 * sizeof(int16_t));
185 memcpy(dst0, dst, 4 * 4 * 4);
186 memcpy(dst1, dst, 4 * 4 * 4);
187 memcpy(subcoef0, coef, 4 * 4 * 4 * sizeof(int16_t));
188 memcpy(subcoef1, coef, 4 * 4 * 4 * sizeof(int16_t));
189 call_ref(dst0, subcoef0, stride);
190 call_new(dst1, subcoef1, stride);
191 if (memcmp(dst0, dst1, 4 * 4 * 4) ||
192 memcmp(subcoef0, subcoef1, 4 * 4 * 4 * sizeof(int16_t)))
194 bench_new(dst1, subcoef1, stride);
200 static void check_luma_dc_wht(void)
202 LOCAL_ALIGNED_16(int16_t, dc, [4 * 4]);
203 LOCAL_ALIGNED_16(int16_t, dc0, [4 * 4]);
204 LOCAL_ALIGNED_16(int16_t, dc1, [4 * 4]);
205 int16_t block[4][4][16];
206 LOCAL_ALIGNED_16(int16_t, block0, [4], [4][16]);
207 LOCAL_ALIGNED_16(int16_t, block1, [4], [4][16]);
211 declare_func_emms(AV_CPU_FLAG_MMX, void, int16_t block[4][4][16], int16_t dc[16]);
215 for (blocky = 0; blocky < 4; blocky++) {
216 for (blockx = 0; blockx < 4; blockx++) {
217 uint8_t src[16], dst[16];
218 randomize_buffers(src, dst, 4, block[blocky][blockx]);
220 dct4x4(block[blocky][blockx]);
221 dc[blocky * 4 + blockx] = block[blocky][blockx][0];
222 block[blocky][blockx][0] = rnd();
227 for (dc_only = 0; dc_only <= 1; dc_only++) {
228 void (*idct)(int16_t [4][4][16], int16_t [16]) = dc_only ? d.vp8_luma_dc_wht_dc : d.vp8_luma_dc_wht;
230 if (check_func(idct, "vp8_luma_dc_wht%s", dc_only ? "_dc" : "")) {
232 memset(dc0, 0, 16 * sizeof(int16_t));
235 memcpy(dc0, dc, 16 * sizeof(int16_t));
237 memcpy(dc1, dc0, 16 * sizeof(int16_t));
238 memcpy(block0, block, 4 * 4 * 16 * sizeof(int16_t));
239 memcpy(block1, block, 4 * 4 * 16 * sizeof(int16_t));
240 call_ref(block0, dc0);
241 call_new(block1, dc1);
242 if (memcmp(block0, block1, 4 * 4 * 16 * sizeof(int16_t)) ||
243 memcmp(dc0, dc1, 16 * sizeof(int16_t)))
245 bench_new(block1, dc1);
250 #define SRC_BUF_STRIDE 32
251 #define SRC_BUF_SIZE (((size << (size < 16)) + 5) * SRC_BUF_STRIDE)
252 // The mc subpixel interpolation filter needs the 2 previous pixels in either
253 // direction, the +1 is to make sure the actual load addresses always are
255 #define src (buf + 2 * SRC_BUF_STRIDE + 2 + 1)
257 #undef randomize_buffers
258 #define randomize_buffers() \
261 for (k = 0; k < SRC_BUF_SIZE; k += 4) { \
262 AV_WN32A(buf + k, rnd()); \
266 static void check_mc(void)
268 LOCAL_ALIGNED_16(uint8_t, buf, [32 * 32]);
269 LOCAL_ALIGNED_16(uint8_t, dst0, [16 * 16]);
270 LOCAL_ALIGNED_16(uint8_t, dst1, [16 * 16]);
273 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *, ptrdiff_t, uint8_t *, ptrdiff_t, int, int, int);
277 for (type = 0; type < 2; type++) {
278 vp8_mc_func (*tab)[3][3] = type ? d.put_vp8_bilinear_pixels_tab : d.put_vp8_epel_pixels_tab;
279 for (k = 1; k < 8; k++) {
281 int size = 16 >> hsize;
282 int height = (size << 1) >> (k % 3);
283 for (dy = 0; dy < 3; dy++) {
284 for (dx = 0; dx < 3; dx++) {
288 static const char *dx_names[] = { "", "h4", "h6" };
289 static const char *dy_names[] = { "", "v4", "v6" };
290 snprintf(str, sizeof(str), "epel%d_%s%s", size, dx_names[dx], dy_names[dy]);
292 snprintf(str, sizeof(str), "bilin%d_%s%s", size, dx ? "h" : "", dy ? "v" : "");
295 snprintf(str, sizeof(str), "pixels%d", size);
297 if (check_func(tab[hsize][dy][dx], "vp8_put_%s", str)) {
301 mx = dx == 2 ? 2 + 2 * (rnd() % 3) : dx == 1 ? 1 + 2 * (rnd() % 4) : 0;
302 my = dy == 2 ? 2 + 2 * (rnd() % 3) : dy == 1 ? 1 + 2 * (rnd() % 4) : 0;
304 mx = dx ? 1 + (rnd() % 7) : 0;
305 my = dy ? 1 + (rnd() % 7) : 0;
308 for (i = -2; i <= 3; i++) {
309 int val = (i == -1 || i == 2) ? 0 : 0xff;
310 // Set pixels in the first row and column to the maximum pattern,
311 // to test for potential overflows in the filter.
313 src[i * SRC_BUF_STRIDE] = val;
315 call_ref(dst0, size, src, SRC_BUF_STRIDE, height, mx, my);
316 call_new(dst1, size, src, SRC_BUF_STRIDE, height, mx, my);
317 if (memcmp(dst0, dst1, size * height))
319 bench_new(dst1, size, src, SRC_BUF_STRIDE, height, mx, my);
327 #undef randomize_buffers
329 #define setpx(a, b, c) buf[(a) + (b) * jstride] = av_clip_uint8(c)
330 // Set the pixel to c +/- [0,d]
331 #define setdx(a, b, c, d) setpx(a, b, c - (d) + (rnd() % ((d) * 2 + 1)))
332 // Set the pixel to c +/- [d,d+e] (making sure it won't be clipped)
333 #define setdx2(a, b, o, c, d, e) setpx(a, b, o = c + ((d) + (rnd() % (e))) * (c >= 128 ? -1 : 1))
335 static void randomize_loopfilter_buffers(int lineoff, int str,
336 int dir, int flim_E, int flim_I,
337 int hev_thresh, uint8_t *buf,
340 uint32_t mask = 0xff;
341 int off = dir ? lineoff : lineoff * str;
342 int istride = dir ? 1 : str;
343 int jstride = dir ? str : 1;
345 for (i = 0; i < 8; i += 2) {
346 // Row 0 will trigger hev for q0/q1, row 2 will trigger hev for p0/p1,
347 // rows 4 and 6 will not trigger hev.
348 // force_hev 1 will make sure all rows trigger hev, while force_hev -1
349 // makes none of them trigger it.
350 int idx = off + i * istride, p2, p1, p0, q0, q1, q2;
351 setpx(idx, 0, q0 = rnd() & mask);
352 if (i == 0 && force_hev >= 0 || force_hev > 0)
353 setdx2(idx, 1, q1, q0, hev_thresh + 1, flim_I - hev_thresh - 1);
355 setdx(idx, 1, q1 = q0, hev_thresh);
356 setdx(idx, 2, q2 = q1, flim_I);
357 setdx(idx, 3, q2, flim_I);
358 setdx(idx, -1, p0 = q0, flim_E >> 2);
359 if (i == 2 && force_hev >= 0 || force_hev > 0)
360 setdx2(idx, -2, p1, p0, hev_thresh + 1, flim_I - hev_thresh - 1);
362 setdx(idx, -2, p1 = p0, hev_thresh);
363 setdx(idx, -3, p2 = p1, flim_I);
364 setdx(idx, -4, p2, flim_I);
368 // Fill the buffer with random pixels
369 static void fill_loopfilter_buffers(uint8_t *buf, ptrdiff_t stride, int w, int h)
372 for (y = 0; y < h; y++)
373 for (x = 0; x < w; x++)
374 buf[y * stride + x] = rnd() & 0xff;
377 #define randomize_buffers(buf, lineoff, str, force_hev) \
378 randomize_loopfilter_buffers(lineoff, str, dir, flim_E, flim_I, hev_thresh, buf, force_hev)
380 static void check_loopfilter_16y(void)
382 LOCAL_ALIGNED_16(uint8_t, base0, [32 + 16 * 16]);
383 LOCAL_ALIGNED_16(uint8_t, base1, [32 + 16 * 16]);
385 int dir, edge, force_hev;
386 int flim_E = 20, flim_I = 10, hev_thresh = 7;
387 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *, ptrdiff_t, int, int, int);
391 for (dir = 0; dir < 2; dir++) {
392 int midoff = dir ? 4 * 16 : 4;
393 int midoff_aligned = dir ? 4 * 16 : 16;
394 uint8_t *buf0 = base0 + midoff_aligned;
395 uint8_t *buf1 = base1 + midoff_aligned;
396 for (edge = 0; edge < 2; edge++) {
397 void (*func)(uint8_t *, ptrdiff_t, int, int, int) = NULL;
398 switch (dir << 1 | edge) {
399 case (0 << 1) | 0: func = d.vp8_h_loop_filter16y; break;
400 case (1 << 1) | 0: func = d.vp8_v_loop_filter16y; break;
401 case (0 << 1) | 1: func = d.vp8_h_loop_filter16y_inner; break;
402 case (1 << 1) | 1: func = d.vp8_v_loop_filter16y_inner; break;
404 if (check_func(func, "vp8_loop_filter16y%s_%s", edge ? "_inner" : "", dir ? "v" : "h")) {
405 for (force_hev = -1; force_hev <= 1; force_hev++) {
406 fill_loopfilter_buffers(buf0 - midoff, 16, 16, 16);
407 randomize_buffers(buf0, 0, 16, force_hev);
408 randomize_buffers(buf0, 8, 16, force_hev);
409 memcpy(buf1 - midoff, buf0 - midoff, 16 * 16);
410 call_ref(buf0, 16, flim_E, flim_I, hev_thresh);
411 call_new(buf1, 16, flim_E, flim_I, hev_thresh);
412 if (memcmp(buf0 - midoff, buf1 - midoff, 16 * 16))
415 fill_loopfilter_buffers(buf0 - midoff, 16, 16, 16);
416 randomize_buffers(buf0, 0, 16, 0);
417 randomize_buffers(buf0, 8, 16, 0);
418 bench_new(buf0, 16, flim_E, flim_I, hev_thresh);
424 static void check_loopfilter_8uv(void)
426 LOCAL_ALIGNED_16(uint8_t, base0u, [32 + 16 * 16]);
427 LOCAL_ALIGNED_16(uint8_t, base0v, [32 + 16 * 16]);
428 LOCAL_ALIGNED_16(uint8_t, base1u, [32 + 16 * 16]);
429 LOCAL_ALIGNED_16(uint8_t, base1v, [32 + 16 * 16]);
431 int dir, edge, force_hev;
432 int flim_E = 20, flim_I = 10, hev_thresh = 7;
433 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *, uint8_t *, ptrdiff_t, int, int, int);
437 for (dir = 0; dir < 2; dir++) {
438 int midoff = dir ? 4 * 16 : 4;
439 int midoff_aligned = dir ? 4 * 16 : 16;
440 uint8_t *buf0u = base0u + midoff_aligned;
441 uint8_t *buf0v = base0v + midoff_aligned;
442 uint8_t *buf1u = base1u + midoff_aligned;
443 uint8_t *buf1v = base1v + midoff_aligned;
444 for (edge = 0; edge < 2; edge++) {
445 void (*func)(uint8_t *, uint8_t *, ptrdiff_t, int, int, int) = NULL;
446 switch (dir << 1 | edge) {
447 case (0 << 1) | 0: func = d.vp8_h_loop_filter8uv; break;
448 case (1 << 1) | 0: func = d.vp8_v_loop_filter8uv; break;
449 case (0 << 1) | 1: func = d.vp8_h_loop_filter8uv_inner; break;
450 case (1 << 1) | 1: func = d.vp8_v_loop_filter8uv_inner; break;
452 if (check_func(func, "vp8_loop_filter8uv%s_%s", edge ? "_inner" : "", dir ? "v" : "h")) {
453 for (force_hev = -1; force_hev <= 1; force_hev++) {
454 fill_loopfilter_buffers(buf0u - midoff, 16, 16, 16);
455 fill_loopfilter_buffers(buf0v - midoff, 16, 16, 16);
456 randomize_buffers(buf0u, 0, 16, force_hev);
457 randomize_buffers(buf0v, 0, 16, force_hev);
458 memcpy(buf1u - midoff, buf0u - midoff, 16 * 16);
459 memcpy(buf1v - midoff, buf0v - midoff, 16 * 16);
461 call_ref(buf0u, buf0v, 16, flim_E, flim_I, hev_thresh);
462 call_new(buf1u, buf1v, 16, flim_E, flim_I, hev_thresh);
463 if (memcmp(buf0u - midoff, buf1u - midoff, 16 * 16) ||
464 memcmp(buf0v - midoff, buf1v - midoff, 16 * 16))
467 fill_loopfilter_buffers(buf0u - midoff, 16, 16, 16);
468 fill_loopfilter_buffers(buf0v - midoff, 16, 16, 16);
469 randomize_buffers(buf0u, 0, 16, 0);
470 randomize_buffers(buf0v, 0, 16, 0);
471 bench_new(buf0u, buf0v, 16, flim_E, flim_I, hev_thresh);
477 static void check_loopfilter_simple(void)
479 LOCAL_ALIGNED_16(uint8_t, base0, [32 + 16 * 16]);
480 LOCAL_ALIGNED_16(uint8_t, base1, [32 + 16 * 16]);
483 int flim_E = 20, flim_I = 30, hev_thresh = 0;
484 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *, ptrdiff_t, int);
488 for (dir = 0; dir < 2; dir++) {
489 int midoff = dir ? 4 * 16 : 4;
490 int midoff_aligned = dir ? 4 * 16 : 16;
491 uint8_t *buf0 = base0 + midoff_aligned;
492 uint8_t *buf1 = base1 + midoff_aligned;
493 void (*func)(uint8_t *, ptrdiff_t, int) = dir ? d.vp8_v_loop_filter_simple : d.vp8_h_loop_filter_simple;
494 if (check_func(func, "vp8_loop_filter_simple_%s", dir ? "v" : "h")) {
495 fill_loopfilter_buffers(buf0 - midoff, 16, 16, 16);
496 randomize_buffers(buf0, 0, 16, -1);
497 randomize_buffers(buf0, 8, 16, -1);
498 memcpy(buf1 - midoff, buf0 - midoff, 16 * 16);
499 call_ref(buf0, 16, flim_E);
500 call_new(buf1, 16, flim_E);
501 if (memcmp(buf0 - midoff, buf1 - midoff, 16 * 16))
503 bench_new(buf0, 16, flim_E);
508 void checkasm_check_vp8dsp(void)
516 check_loopfilter_16y();
517 check_loopfilter_8uv();
518 check_loopfilter_simple();
519 report("loopfilter");